What If Pharma Already Has Too Many Drugs?

Formation Bio has a slightly strange way of thinking about AI in drug discovery.

This New York company started life as TrialSpark in 2016, founded by Benjamine Liu and Linhao Zhang. Liu was a computational biologist at Oxford and, like others, believed AI could help discover better drugs. But when he tried pitching some Alzheimer’s ideas to pharma companies, he says they told him they already had more drugs than they could afford to develop.

This became the company thesis. TrialSpark first built software for clinical trials, then became a CRO that ran them. Eventually it changed its name to Formation Bio and started buying and licensing drugs itself. Today, Formation has raised about $615 million at a $1.8 billion valuation, backed by investors including Andreessen Horowitz, Sequoia, Thrive, John Doerr and Sam Altman.

This week, its new R&D chief Michael Ehlers, previously head of R&D at Biogen and a senior research leader at Pfizer, described what he thinks is becoming a bigger problem for pharma. AI and modern drug discovery are producing an “excess of clinical-stage molecules,” while clinical development remains the bottleneck.

Formation’s answer is to find promising drugs that other companies have shelved, deprioritized or cannot afford to keep developing, then use AI, clinical data and experienced drug developers to decide which ones deserve another shot.

There is something appealing about that model because pharma shelves drugs for all kinds of reasons. A company changes strategy or a new CEO comes in or two companies merge, and of course, if money gets tight. So a perfectly decent drug candidate can end up sitting on a shelf.

But there is an old economics problem behind all of this.

Think about buying a used car. The seller knows whether the transmission makes a strange noise every morning. You don’t. That information gap makes buyers suspicious of every used car, including the good ones.

Drug licensing has the same problem.

If Pfizer spent seven years studying a drug and then decided it wasn’t worth another $100 million, Pfizer probably knows a lot more about that drug than the company trying to buy it. Some abandoned drugs really are hidden gems.

This is where Formation’s AI becomes more interesting than another molecule generator. The job is partly to close that information gap. Its systems can pull together clinical results, regulatory history, genetics, competing programs and trial data, then help rank which assets and indications look most promising.

Formation is even pursuing a strategy called “Known in New,” where it starts with biology already validated in humans and looks for new diseases where that mechanism might work. It may also tell us something unexpected about where AI is taking pharma.

We often imagine AI pushing scientists toward strange new targets and unexplored biology. But a model trained to improve the odds of clinical success may learn the opposite lesson. Human validation, regulatory precedent, and known mechanisms are all valuable.

One of AI’s first big contributions to pharma could therefore be making the industry better at avoiding biological risk.

Perhaps this is the more interesting future for AI drug discovery. The biggest winners may not be the companies generating the most novel molecules. They may be the ones that become unusually good at telling a hidden gem from a lemon before anyone else can.

The FDA is trying to cut months off the path from a drug candidate to the first human trial. Under a new initiative called Operation TrialBlazer, the agency is proposing an expedited IND pilot, allowing rolling submissions and closer collaboration with qualified research institutions before Phase 1. The FDA also says companies often submit more manufacturing data than they actually need at this stage, and its new phase-specific CMC approach could save 6 to 12 months. If AI keeps producing candidates faster, regulators are now starting to work on the other side of the equation: getting the good ones into humans faster too. [Link]

Former OpenAI chief product officer Kevin Weil is reportedly raising $150 million for a new AI science startup at a valuation of at least $750 million. Details are still limited, but the company is expected to use AI for scientific discovery and, importantly, generate scientific data that can be used to train the models themselves. Weil is joining a growing group of top AI researchers moving from general-purpose models into science. The interesting shift is that the next AI race may depend less on who has the biggest model and more on who can build the experimental machines that continuously feed those models new data. [Link]

The UK is putting up to £20 million behind a new hub for human-based models that could replace some animal testing in drug development. Researchers will develop more standardized human in-vitro systems, including organoids, with the Cambridge effort initially looking at areas such as inflammatory bowel disease before expanding into cancer and neurological disease. The AI connection is indirect but important. Better models need better experimental feedback, and miniature human tissues could give drug-discovery systems something much closer to human biology to learn from than another mouse experiment. [Link]